Robotics developers can investigate fast local control and combine action prediction with a separate reasoning model for difficult planning. Core capabilities include Seven-billion-parameter action model, Robot action prediction, Joint video-action training, and Embodiment-specific fine-tuning.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Released model weights, Open fine-tuning recipe, Low-latency policy inference, and Hybrid planning-and-control experiments.
